What AI Is Constitutionally Forbidden from Doing

Within Providence's constitutional architecture, artificial intelligence is forbidden from performing functions that substitute for human judgment about meaning, ethics, and governance, and from performing functions whose effects on participants cannot be made fully transparent and contestable.

AI cannot determine trustworthiness. It can assist in processing and presenting attestations. It cannot make the judgment that a participant is or is not trustworthy. That judgment is human and relational, and delegating it to an AI system — however sophisticated — would both undermine the constitutional principle that trust is relational and create a technical oracle whose decisions would, in practice, function as governance authority regardless of what the formal governance structure said.

AI cannot make binding governance decisions. It can assist in governance processes. It cannot decide outcomes. The constitutional principle of participatory legitimacy requires that governance decisions be made through processes in which participants can genuinely participate, which means processes in which the decision-making is comprehensible to participants, not delegated to systems whose operation is opaque even to experts.

AI cannot operate on participant data without specific, ongoing, fully informed consent. The constitutional principle of data sovereignty requires that participants control how information about them is used. An AI system that continuously processes participant data on the basis of consent given at the time of joining the network — without ongoing transparency about what the system is doing and ongoing ability to withdraw consent without losing access to core network functions — is not operating within the constitutional principles regardless of how benign its operations are.

AI cannot be deployed in ways that are not fully auditable by participants and their designated stewards. Opacity is a constitutional violation regardless of whether the opaque system is producing good outcomes. The constitutional principle of transparency requires that participants can understand what systems are doing and why, and can contest those operations when they believe the operations are inconsistent with the constitutional principles.

The AI alignment literature — the body of work concerned with how AI systems can be designed to pursue the objectives that their designers actually intend rather than the objectives that are technically specified — is directly relevant to Providence's AI governance design. Stuart Russell's Human Compatible (2019) provides the most accessible account of the alignment problem and the reasons it is a genuine problem rather than merely a theoretical concern. The work of the Center for Human-Compatible AI at Berkeley, the work of the Machine Intelligence Research Institute, and the work of the AI safety team at DeepMind all bear on the specific technical question of how AI constraints can be made robust.

The political philosophy of algorithmic governance has developed rapidly in the past decade. Frank Pasquale's The Black Box Society (2015) and New Laws of Robotics (2020), Virginia Eubanks' Automating Inequality (2018), Safiya Umoja Noble's Algorithms of Oppression (2018), and Kate Crawford's Atlas of AI (2021) collectively provide an extensive analysis of how AI systems embedded in existing power structures reproduce and amplify those structures' inequities. Providence's AI governance design must engage this literature seriously: the constitutional commitment to human dignity and participatory legitimacy is not satisfied by AI systems that are well-intentioned but whose operations reproduce existing hierarchies of disadvantage.

The emerging literature on AI auditing and accountability — including the work of the Algorithmic Justice League, the AI Now Institute, and the Partnership on AI — provides practitioner-oriented frameworks for how AI systems can be made accountable in institutional settings. The work is imperfect and the tools are incomplete, but they represent the current state of the art for the specific problem of making AI systems governable within institutions that are genuinely committed to doing so.

The deepest open question about AI governance within Providence concerns what happens as AI capabilities advance beyond the current state of the art. The constitutional prohibitions in this chapter are drawn in terms of what AI is doing — making governance decisions, determining trustworthiness, operating without transparency — rather than in terms of specific technical capabilities. This means the prohibitions should remain relevant as capabilities advance: an AI system with dramatically more sophisticated capabilities would still be prohibited from making binding governance decisions, regardless of how much better its decisions might be than the human decisions it would replace. But the pressure to cross the prohibition line will grow as AI capabilities advance, and the governance mechanisms for holding the line will face increasingly sophisticated challenges. How those mechanisms are maintained under that pressure is a question that the current state of the art does not yet answer.